HackerNews Digest

Claude Code “routines” automate tasks via scheduled, API‑triggered, or GitHub‑event triggers (research preview; behavior and limits may change). Users define a routine by naming it, writing a self‑contained prompt, selecting a model, and linking one or more GitHub repositories (cloned at run start; Claude creates `claude/`‑prefixed branches; unrestricted push permission may be required). Environment settings include network access, secret variables, and a setup script for dependencies. Triggers are added through the UI, CLI, or desktop app: - **Schedule** – recurring cadence (hourly, daily, etc.). - **API** – HTTP POST to `https://api.anthropic.com/v1/claude_code/routines//fire` with bearer token and required headers (`anthropic-beta`, `anthropic-version`, `Content-Type`). Response returns a session ID and URL. - **GitHub** – events such as `pull_request.opened`/`closed`; requires installing the Claude GitHub App and can be filtered by branch, fork status, draft flag, or labels. Runs generate separate sessions viewable in the Claude Code interface; users can run immediately, pause schedules, edit prompts, repositories, environments, or delete routines. Past sessions persist after deletion. Limits and API shapes are subject to change under dated beta headers.

- Chicago music fan Aadam Jacobs has recorded live shows since the 1980s, accumulating over 10 000 cassette tapes. - To prevent tape degradation, volunteers from the nonprofit Internet Archive began digitizing the collection; about 2 500 tapes are now online. - The digital archive includes rare early performances such as a 1989 Nirvana set (pre‑“Smells Like Teen Spirit”), plus previously unknown recordings from Sonic Youth, R.E.M., Phish, Liz Phair, Pavement, Neutral Milk Hotel, and numerous punk groups. - Volunteer audio engineer Brian Emerick regularly collects boxes of tapes from Jacobs, plays them on vintage cassette decks, and converts them to digital files. - Additional volunteers clean, organize, and label the files, often researching obscure song titles for forgotten bands. - Despite Jacobs’ original low‑grade equipment, the restoration process has significantly improved audio quality, exemplified by a 1988 Tracy Chapman recording now available on the Archive.

Flock Safety sells AI‑driven “vehicle‑fingerprint” cameras that record license plates plus visual traits (color, make, model, dents, bumper‑sticker placement, roof racks, wheel type) and generate searchable profiles. The platform adds “convoy analysis” to link vehicles that frequently travel together and stores all data in a nationwide law‑enforcement network accessible without a warrant. As of 2025, over 100,000 cameras are deployed, with more than 3,000 agencies using the system; a public map shows roughly half of the units. Legal challenges cite Fourth‑Amendment violations, with a 2024 trial court labeling the network a city‑wide dragnet equivalent to GPS tracking. Documented abuses include a Kansas chief’s 228 unauthorized stings of an ex‑girlfriend and a journalist’s 300‑mile route being logged by ~50 cameras. Demographic analyses reveal disproportionate stops of Black drivers (84 % in Oak Park). Flock expands via partnerships with HOAs, retailers (e.g., Lowe’s) and employers, sharing footage with police. The system is part of a broader trend toward mass AI surveillance, intersecting with tools such as Palantir’s data‑integration platforms, raising significant privacy and civil‑rights concerns.

The post contrasts current AI hype with the disciplined practice of reverse engineering. It recounts how Bulgaria’s Pravetz (Правец) computers were built in the 1970s‑80s as Apple II clones: engineers at Sofia’s Institute of Technical Cybernetics copied the 6502‑based design, replaced the ROM and case, added Cyrillic uppercase character encoding, and later integrated Z80 CPUs and CP/M support. By the mid‑1980s Pravetz accounted for ~40 % of personal computers in COMECON, illustrating how detailed schematic replication enabled a national computing base without proprietary access. The article then describes the ISCAS‑85 benchmark suite, released in 1985 as gate‑level netlists without functional documentation. In 1999, Hayes, Hansen, and Yalcin reverse‑engineered all ten circuits, revealing real hardware functions (e.g., a 27‑channel interrupt controller, an 8‑bit ALU, a 16×16 multiplier). Their work provided functional specifications that improve test generation, synthesis verification, and hierarchical analysis. Finally, the author links reverse engineering, model‑based diagnosis, and synthesis as mathematically dual problems—each solving for a circuit’s behavior given different fixed variables—arguing that careful analysis yields lasting understanding unlike opaque, resource‑intensive AI development.

The Orange Pi 6 Plus uses the CIX P1 (CD8180/CD8160) SoC with a 12‑core big.LITTLE CPU (4 × Cortex‑A520 @ 1.8 GHz, 8 × Cortex‑A720 @ 2.6 GHz), Mali‑G720/Immortalis GPU, three‑core Zhouyi NPU, 16 GiB RAM (≈14 GiB usable), dual Realtek RTL8126 5 GbE, RTL8852BE Wi‑Fi/Bluetooth, and standard USB. The board ships with CIX UEFI 1.3. The author built a custom Debian 13 (Trixie) image via a fork of orangepi‑build, removing Ubuntu‑specific packages, adding reproducible kernel 6.6.89‑cix, and integrating GPU/NPU firmware and development tools. Initial boot failures were due to an incorrect EFI stub and a partition‑resize script; patches corrected GRUB entries and the resize helper. NVMe boot was achieved by migrating EFI, root and swap partitions from the SD card. GPU support required installing vendor userspace (cix‑gpu‑umd, libglvnd, libdrm, libmesa) and rebinding the Mali driver; Vulkan then reported a Mali‑G720‑Immortalis device. NPU kernels detected three cores, but userspace packages were fragmented and needed manual validation. Inference benchmarks (llama.cpp, ik_llama, PowerInfer) showed the Qwen 3.5 4 B model on Vulkan as the most stable, delivering ~8 tokens /s with acceptable memory usage; larger models caused failures or excessive RAM consumption. Vulkan descriptor‑set limits required micro‑batch tuning for reliable operation. Overall, the board is functional for edge AI within a limited model/runtime envelope.

The analysis of Picasso’s Guernica combines infrared reflectography, structural assessment, and conservation history. Reflectography shows early compositional revisions: the soldier’s head was initially inverted and later rotated, revealing numerous eyes, while hidden eyes appear in the bull’s head beneath the white layer. The canvas suffers severe support damage from repeated nailing and removal of its original unstressed frame, leaving torn, holed edges. Extensive handling created cracks and fissures; fissures are more numerous, whereas deeper cracks affect multiple layers (preparation, paint, protective) and correlate with thicker impasto. A 1957 MoMA restoration introduced a wax‑resin consolidation applied from the reverse, now detectable as surface wax and under UV fluorescence. In 1974 a red acrylic spray attack was removed, leaving only microscopic red residues in some fissures.

None

The article critiques common 5NF teaching examples as artificial and argues that proper design starts with a clear logical model derived from business requirements. It surveys three typical sources—the Wikipedia salesman/brand/product‑type case, the “ice‑cream” brand‑flavour‑friend scenario, and the “musicians” concert‑instrument‑musician example—and shows their shortcomings. From these, two generic patterns emerge: * **AB‑BC‑AC triangle** – three many‑to‑many links connecting three anchors (e.g., brands, flavours, friends). Implemented with three two‑column junction tables. * **ABC + D star** – a fourth anchor (Performance) links three primary anchors; links are 1:N and can be stored in a single table with either a composite primary key (concert_id, musician_id, instrument_id) or a synthetic ID plus a uniqueness constraint. The article demonstrates how to choose composite versus synthetic keys based on uniqueness requirements and extends the models (e.g., adding specific preferences or musician skills). It concludes that 5NF decomposition is unnecessary when the logical model is built correctly; the two patterns yield fully normalized schemas without redundancy or anomalies.

Clojure’s persistent vectors are immutable tree‑based structures that achieve effectively constant‑time (`O(log₃₂ n)`) appends, updates, lookups and sub‑vector operations. They replace mutable array‑list semantics with a balanced, ordered tree whose interior nodes have a fixed branching factor (32 in Clojure), and leaf nodes store actual elements. Persistence is obtained by *path copying*: when inserting, updating, or removing an element, only the nodes along the path to the affected leaf are duplicated, while the remaining sub‑trees are shared between the old and new vectors. **Updates** traverse the tree to the target leaf, copy each node on the path, replace the value, and return the new root. **Appends** are similar but must handle three cases: (1) space in the rightmost leaf, (2) need to create intermediate nodes when the leaf is full, and (3) root overflow, which occurs when the size is a power of the branching factor and requires a new root node. **Popping** mirrors append logic with three cases: (1) leaf still has elements after removal, (2) leaf becomes empty and is pruned, propagating null pointers upward, and (3) root contains a single child and is collapsed. Because the tree depth is shallow (max 6 levels for ≤ 1 billion elements), these operations are treated as “effectively O(1)” despite their logarithmic nature. Subsequent posts will cover tails, transients, and sub‑vector mechanics.

The guide documents installing OpenBSD‑current on Japanese‑model Pomera DM250/DM250X/DM250XY devices using a custom kernel and U‑Boot image. It warns of risks: full battery drain can render the device unbootable, and loss of U‑Boot may require USB‑C recovery or opening the device. A full eMMC backup via tools from EKESETE.net is recommended. The factory U‑Boot boots a recovery Linux kernel when Right Shift + Left Alt are held at power‑on, which runs an _sdboot.sh script from an SD card. The script backs up firmware, replaces U‑Boot with the provided image, and then reboots into the new EFI‑enabled U‑Boot. Preparing the SD card involves GPT partitioning with a ≥100 MiB EFI partition, copying BOOTARM.EFI, uboot.img, the _sdboot.sh script, OpenBSD armv7 snapshot sets, bwfm Wi‑Fi firmware, and custom bsd/bsd.rd images. Installation boots via the new U‑Boot, uses the “whole” disk option to create an MBR with an EFI offset, and disables the upstream reorder_kernel script. Post‑install steps include restoring Wi‑Fi firmware (nvram_ap6212a.txt) and optional logo.bmp. Recovery methods cover MaskROM/loader mode USB mass‑storage access with xrock and a restore script, as well as accessing the U‑Boot prompt or UART if needed.